Palmist: a Tool to Log Palm System Activity (Under the Direction of SURENDAR CHANDRA)
Total Page：16
File Type：pdf, Size：1020Kb
RAGHUNATH GANNAMARAJU Palmist: A Tool to Log Palm System Activity (Under the direction of SURENDAR CHANDRA) With the proliferation in the usage scenarios of mobile handheld devices, understanding typical client usage patterns is fundamental to developing new policies to improve the usability of these devices. In this thesis, we describe a Palm system call logging tool called Palmist. Palmist allows the practitioner to selectively collect statistics such as the system call invoked, application that invoked the system call, the time of the call and the system call arguments. Our logging mechanism consumes about 20 bytes of memory on the PDA to store the log record. The logging mechanism adds a latency of about 10 msec to collect the log. The mechanism has limitations in collecting logs for system calls that are needed by the collection mechanism itself. Our logging mechanism works for about 88% (735 of 834 relevant system calls) of the Palm OS 3.5 system calls. Our system can be utilized by system developers to customize their application behavior to optimize system parameters such as energy consumption and ease of use. INDEX WORDS: Palm OS, Mobile User Access Pattern, Thesis (academic) PALMIST: A TOOL TO LOG PALM SYSTEM ACTIVITY by RAGHUNATH GANNAMARAJU B.E., Birla Institute of Technology, India, 1997 A Thesis Submitted to the Graduate Faculty of The University of Georgia in Partial Fulfillment of the Requirements for the Degree MASTER OF SCIENCE ATHENS, GEORGIA 2001 c 2001 Raghunath Gannamaraju All Rights Reserved PALMIST: A TOOL TO LOG PALM SYSTEM ACTIVITY by RAGHUNATH GANNAMARAJU Approved: Major Professor: Surendar Chandra Committee: David Lowenthal Eileen T. Kraemer Electronic Version Approved: Gordhan L. Patel Dean of the Graduate School The University of Georgia December 2001 ACKNOWLEDGMENTS I would like to thank Dr. Surendar Chandra for all the help and guidance he has provided during the course of this thesis. I am grateful for his help during some of the debugging sessions. I would also like to thank Dr. David Lowenthal and Dr. Eileen Kraemer for consenting to serve on my committee. iv TABLE OF CONTENTS Page ACKNOWLEDGMENTS . iv LIST OF TABLES . vii LIST OF FIGURES . viii CHAPTER 1 INTRODUCTION . 1 2 BACKGROUND . 4 2.1 TRAP HANDLING . 4 2.2 FEATURE MANAGER . 5 2.3 PALM DATABASES . 6 2.4 HOTSYNC . 6 3 SYSTEM ARCHITECTURE . 7 3.1 OBJECTIVES . 7 3.2 PALMIST TRAP HANDLER INSTALLATION MECHANISM . 7 3.3 PALMIST TRAP HANDLERS . 9 3.4 DOWNLOADING PALM’S ACTIVITY LOG ONTO THE DESKTOP . 14 3.5 EXPERIMENTAL SETUP . 14 4 PERFORMANCE RESULTS . 15 4.1 LOG COLLECTION OVERHEAD . 15 4.2 PALMIST USAGE EXAMPLES . 20 v vi 5 RELATED WORK . 27 6 CONCLUSIONS AND FUTURE DIRECTIONS . 29 BIBLIOGRAPHY . 31 APPENDIX A PSEUDO CODE FOR INSTALLING PALMIST TRAP FUNCTION (SYSTRAP 0XA00E) . 33 B PSEUDO CODE FOR THE LIBRARY INITIALIZATION FUNCTION . 34 C PSEUDO CODE FOR A TYPICAL SYSTEM CALL TRAP HANDLER . 35 D PSUEDO CODE FOR COMMON LIBRARY . 37 E PSEUDO CODE FOR THE CONDUIT TO DOWNLOAD LOGS TO THE DESKTOP 40 LIST OF TABLES 3.1 Distribution of the number of parameters to Palm OS 3.5 system calls logged with Palmist . 10 3.2 Palm event latency (rounded down) . 12 4.1 Palmist overhead per system call . 17 4.2 Relative overhead for Palmist internal system calls . 17 4.3 Time taken for typical operation using Address Book application . 18 4.4 Time taken for typical operation using Calculator application . 19 4.5 Time taken for typical operation using Date Book application . 19 4.6 Pinemark Benchmarks . 23 vii LIST OF FIGURES 2.1 Palmist patch to system traps . 5 3.1 Palmist mechanism to trap system calls . 8 4.1 “Representative” session . 22 4.2 Text Pinemark benchmark (Note: Calls marked with a * were also popular in the “representative” session) . 25 4.3 System calls invoked with time (Text Pinemark benchmark) . 26 viii CHAPTER 1 INTRODUCTION There is an explosion in the number and variety of handheld devices. Dataquest estimates that 3.55 million handheld devices were delivered worldwide during the first quarter of 2001. Palm OS based [8] PDAs account for 87% of the handheld devices sold [13]. Falling prices and the increasing availability of newer applications is driving the popularity of hand held devices. Hand held devices are constrained in their usability by the available network, bat- tery capacity and display characteristics. The typical client usage pattern has a profound impact on the usability and the efficacy of system management policies. For example, the access pattern of the user significantly affects the battery energy consumption. Depending on the idle duration, the Palm device consumes more energy to shut down and power up than to remain idle [4]. A policy that conserves energy by shutting down the unit after a fixed idle period would not be effective if the user resumes work immediately, forcing a powerup. Also, it has been shown [3] that the batteries perform better under pulsed discharge, rather than a steady discharge. Policies that can change the computation to leverage these device characteristics can be expected to prolong battery life. Under- standing the actual user behavior is fundamental to designing these newer policies. Representative client usage logs are invaluable in answering realistic questions about the system behavior of PDAs. However, little is known of the actual usage behavior of hand held devices. Earlier work [5, 4] used oscilloscopes to capture system activity. Such probes restrict the ability to collect mobile user activity. Also, publicly available Palm 1 2 OS Emulator (POSE) [6] is widely used for debugging and profiling Palm applications. These emulation programs typically run on desktop computers. Hence, their applicability to profile and capture mobile users’ behavior is limited. There is a need to instrument the actual devices to capture the mobile user behavior. The primary goal of this work is to develop tools to capture the system behavior of PDA devices. The logs collected can be used by practitioners to evaluate the performance of their system design. For example, operating system developers for the PDA platform can analyze the efficacy of their energy saving features for a typical user. Application designers can also improve the general usability by improving the accessibility of popular features. In this work, we focus our attention on Palm devices. We describe Palmist, a tool to log system calls for the Palm device. One of the design goals in developing Palmist was to capture enough system activity details to replay them in log play back tools. The Palms were instrumented using the hack [9] mechanism to collect statistics such as the system call invoked, application that invoked the system call, the time of the call and the system call arguments. The system allows the practitioner to selectively log a subset of interesting system calls. The collected information can be periodically synced to a desktop computer for further processing. These instrumented Palms can be distributed to typical target users and their usage patterns collected over a period of time. The trends in the users’ access patterns can be analyzed. One of the challenges in designing Palmist was to constrain the overhead of the addi- tional instrumentation. The overhead introduced by Palmist depends on the amount of system call activity logged. We show that our logging mechanism adds an acceptable amount of overhead in collecting the system call events. On the average, the log records consume about 20 bytes of storage on the Palm device. The Palmist trap handlers add an additional latency of about 10 msec per system call logged. Depending on the application scenario, we show that in the worst case, a fully instrumented Palm can run up to two 3 orders of magnitude slower. However, disabling a few popular calls and using a faster PDA for interactive tasks adds an extra 50% overhead only. The mechanism also has lim- itations in collecting logs for system calls that are needed by the collection mechanism itself. Our logging mechanism works for about 88% (735 of 834 of relevant system calls) of the Palm OS system calls. The remainder of this report is organized as follows: in Chapter 2 we begin by setting the background for our work by briefly describing the Palm OS features used in developing our system. We describe the Palmist tool in detail in Chapter 3. Experimental setup to measure the logging overhead and the evaluation methodologies used in our study are described in Chapter 4. Chapter 5 places our work in the context of other related work in the field. We present our conclusions and outline future research directions in Chapter 6. CHAPTER 2 BACKGROUND In this chapter, we briefly describe several features of Palm OS that were exploited in developing Palmist. In particular, we describe how the Palm OS system handlers can be replaced with our own trap handlers, how the feature manager can be utilized to provide global memory regions, how Palm stores permanent data in databases, as well as the HotSync operation to upload data from the Palm to a desktop computer. 2.1 TRAP HANDLING The Palm OS supports a trap mechanism to service system events (such as “user entered a Graffiti character”, “system timer ticked”, “power-off”, etc.). The system events are ser- viced through a system-wide trap handler vector (illustrated in Figure 2.1). Each system call has a 16-bit Trap ID value associated with it. By default, when an event occurs or a system call is executed, a lookup is done in the trap handler vector and the appropriate Palm OS system call trap handler is executed. Based on their functionality, the system calls can be broadly categorized into various APIs.